dyncom: Implement USAT/SSAT

src/core/arm/dyncom/arm_dyncom_interpreter.cpp

@@ -1275,6 +1275,14 @@ typedef struct _smla_inst {
 	unsigned int Rn;
 } smla_inst;
 
+typedef struct ssat_inst {
+	unsigned int Rn;
+	unsigned int Rd;
+	unsigned int imm5;
+	unsigned int sat_imm;
+	unsigned int shift_type;
+} ssat_inst;
+
 typedef struct umaal_inst {
 	unsigned int Rn;
 	unsigned int Rm;
@@ -2742,7 +2750,24 @@ ARM_INST_PTR INTERPRETER_TRANSLATE(smulw)(unsigned int inst, int index)
 }
 ARM_INST_PTR INTERPRETER_TRANSLATE(smusd)(unsigned int inst, int index)    { UNIMPLEMENTED_INSTRUCTION("SMUSD"); }
 ARM_INST_PTR INTERPRETER_TRANSLATE(srs)(unsigned int inst, int index)      { UNIMPLEMENTED_INSTRUCTION("SRS"); }
-ARM_INST_PTR INTERPRETER_TRANSLATE(ssat)(unsigned int inst, int index)     { UNIMPLEMENTED_INSTRUCTION("SSAT"); }
+ARM_INST_PTR INTERPRETER_TRANSLATE(ssat)(unsigned int inst, int index)
+{
+	arm_inst* const inst_base = (arm_inst*)AllocBuffer(sizeof(arm_inst) + sizeof(ssat_inst));
+	ssat_inst* const inst_cream = (ssat_inst*) inst_base->component;
+
+	inst_base->cond     = BITS(inst, 28, 31);
+	inst_base->idx      = index;
+	inst_base->br       = NON_BRANCH;
+	inst_base->load_r15 = 0;
+
+	inst_cream->Rn = BITS(inst, 0, 3);
+	inst_cream->Rd = BITS(inst, 12, 15);
+	inst_cream->imm5 = BITS(inst, 7, 11);
+	inst_cream->sat_imm = BITS(inst, 16, 20);
+	inst_cream->shift_type = BIT(inst, 6);
+
+	return inst_base;
+}
 ARM_INST_PTR INTERPRETER_TRANSLATE(ssat16)(unsigned int inst, int index)   { UNIMPLEMENTED_INSTRUCTION("SSAT16"); }
 ARM_INST_PTR INTERPRETER_TRANSLATE(ssub8)(unsigned int inst, int index)    { UNIMPLEMENTED_INSTRUCTION("SSUB8"); }
 ARM_INST_PTR INTERPRETER_TRANSLATE(ssub16)(unsigned int inst, int index)
@@ -3344,7 +3369,10 @@ ARM_INST_PTR INTERPRETER_TRANSLATE(usad8)(unsigned int inst, int index)
 {
 	return INTERPRETER_TRANSLATE(usada8)(inst, index);
 }
-ARM_INST_PTR INTERPRETER_TRANSLATE(usat)(unsigned int inst, int index)      { UNIMPLEMENTED_INSTRUCTION("USAT"); }
+ARM_INST_PTR INTERPRETER_TRANSLATE(usat)(unsigned int inst, int index)
+{
+	return INTERPRETER_TRANSLATE(ssat)(inst, index);
+}
 ARM_INST_PTR INTERPRETER_TRANSLATE(usat16)(unsigned int inst, int index)    { UNIMPLEMENTED_INSTRUCTION("USAT16"); }
 ARM_INST_PTR INTERPRETER_TRANSLATE(usub16)(unsigned int inst, int index)    { UNIMPLEMENTED_INSTRUCTION("USUB16"); }
 ARM_INST_PTR INTERPRETER_TRANSLATE(usub8)(unsigned int inst, int index)     { UNIMPLEMENTED_INSTRUCTION("USUB8"); }
@@ -6284,6 +6312,40 @@ unsigned InterpreterMainLoop(ARMul_State* state)
 	SMUSD_INST:
 	SRS_INST:
 	SSAT_INST:
+	{
+		INC_ICOUNTER;
+
+		if (inst_base->cond == 0xE || CondPassed(cpu, inst_base->cond)) {
+			ssat_inst* const inst_cream = (ssat_inst*)inst_base->component;
+			
+			u8 shift_type = inst_cream->shift_type;
+			u8 shift_amount = inst_cream->imm5;
+			u32 rn_val = RN;
+
+			// 32-bit ASR is encoded as an amount of 0.
+			if (shift_type == 1 && shift_amount == 0)
+				shift_amount = 31;
+
+			if (shift_type == 0)
+				rn_val <<= shift_amount;
+			else if (shift_type == 1)
+				rn_val = ((s32)rn_val >> shift_amount);
+
+			bool saturated = false;
+			rn_val = ARMul_SignedSatQ(rn_val, inst_cream->sat_imm, &saturated);
+
+			if (saturated)
+				cpu->Cpsr |= (1 << 27);
+
+			RD = rn_val;
+		}
+
+		cpu->Reg[15] += GET_INST_SIZE(cpu);
+		INC_PC(sizeof(ssat_inst));
+		FETCH_INST;
+		GOTO_NEXT_INST;
+	}
+
 	SSAT16_INST:
 	SSUB8_INST:
 	STC_INST:
@@ -7140,6 +7202,40 @@ unsigned InterpreterMainLoop(ARMul_State* state)
 	}
 
 	USAT_INST:
+	{
+		INC_ICOUNTER;
+
+		if (inst_base->cond == 0xE || CondPassed(cpu, inst_base->cond)) {
+			ssat_inst* const inst_cream = (ssat_inst*)inst_base->component;
+
+			u8 shift_type = inst_cream->shift_type;
+			u8 shift_amount = inst_cream->imm5;
+			u32 rn_val = RN;
+
+			// 32-bit ASR is encoded as an amount of 0.
+			if (shift_type == 1 && shift_amount == 0)
+				shift_amount = 31;
+
+			if (shift_type == 0)
+				rn_val <<= shift_amount;
+			else if (shift_type == 1)
+				rn_val = ((s32)rn_val >> shift_amount);
+
+			bool saturated = false;
+			rn_val = ARMul_UnsignedSatQ(rn_val, inst_cream->sat_imm, &saturated);
+
+			if (saturated)
+				cpu->Cpsr |= (1 << 27);
+
+			RD = rn_val;
+		}
+
+		cpu->Reg[15] += GET_INST_SIZE(cpu);
+		INC_PC(sizeof(ssat_inst));
+		FETCH_INST;
+		GOTO_NEXT_INST;
+	}
+
 	USAT16_INST:
 	USUB16_INST:
 	USUB8_INST:

src/core/arm/interpreter/armsupp.cpp

@@ -578,6 +578,42 @@ u16 ARMul_UnsignedSaturatedSub16(u16 left, u16 right)
     return left - right;
 }
 
+// Signed saturation.
+u32 ARMul_SignedSatQ(s32 value, u8 shift, bool* saturation_occurred)
+{
+    const u32 max = (1 << shift) - 1;
+    const s32 top = (value >> shift);
+
+    if (top > 0) {
+        *saturation_occurred = true;
+        return max;
+    }
+    else if (top < -1) {
+        *saturation_occurred = true;
+        return ~max;
+    }
+
+    *saturation_occurred = false;
+    return (u32)value;
+}
+
+// Unsigned saturation
+u32 ARMul_UnsignedSatQ(s32 value, u8 shift, bool* saturation_occurred)
+{
+    const u32 max = (1 << shift) - 1;
+
+    if (value < 0) {
+        *saturation_occurred = true;
+        return 0;
+    } else if ((u32)value > max) {
+        *saturation_occurred = true;
+        return max;
+    }
+
+    *saturation_occurred = false;
+    return (u32)value;
+}
+
 
 /* This function does the work of generating the addresses used in an
    LDC instruction.  The code here is always post-indexed, it's up to the

src/core/arm/skyeye_common/armdefs.h

@@ -801,6 +801,9 @@ extern u8 ARMul_UnsignedSaturatedSub8(u8, u8);
 extern u16 ARMul_UnsignedSaturatedSub16(u16, u16);
 extern u8 ARMul_UnsignedAbsoluteDifference(u8, u8);
 
+extern u32 ARMul_SignedSatQ(s32, u8, bool*);
+extern u32 ARMul_UnsignedSatQ(s32, u8, bool*);
+
 #define DIFF_LOG 0
 #define SAVE_LOG 0